Abstract:

The EU requirements for "Trustworthy AI" emphasize high-quality data, compliance with legal regulations in data processing and storage, system control, and ensuring AI serves the well-being of users. From the outset, the integration of AI-driven chat assistants in academia should follow these trust requirements. This can be achieved through a framework that balances algorithmic intelligence and generative intelligence to optimize both trustworthy data and conversational quality.

At the heart of the proposed architecture, the operation control component orchestrates the entire system, ensuring high-quality data retrieval, enforcing compliance through guardians, and refining response processing to optimize AI-generated outputs—all within a user-centric interface that facilitates effective interaction. Generative AI is only selectively used where necessary, such as for analyzing or generating text, while the core processes remain anchored in a transparent, rule-based algorithmic framework.

The proposed components serve as conceptual elements to facilitate discussions among academic stakeholders, ensuring a structured and practical AI implementation without requiring deep technical knowledge. Nevertheless, these conceptual components are designed for future implementation as technical modules. Conceptual components include, for example:
- Access & Quota Controller (AQC): Ensures controlled access and responsible AI resource allocation.
- Algorithmic Prompt Composer (APC): Constructs optimized prompts based on institutional, user, and task-relevant data, enhancing response accuracy.
- Qualified Data Retriever (QDR): Retrieves and processes structured data from internal and external databases to support the algorithmic construction of a prompt.
- Vector Document Fetcher (VDF): Retrieves relevant academic knowledge using vector-based search to ground and enrich AI-generated responses.
- Response Data Splitter (RDS): Separates the AI-generated output intended for the user from the metadata used for memory and monitoring.
- Critical Content Guard (CCG): Filters and mitigates biased or inappropriate AI outputs, ensuring institutional compliance.
- Chat Dialogue Memory (CDM): Enables contextual conversation continuity by managing both short-term and long-term interactions.

The interplay of these components results in a framework that serves as a structured approach for embedding trust from the outset in AI systems. By ensuring a seamless interplay between algorithmic processes and generative AI, the proposed architecture actively reinforces "trust-driven AI" by structuring data retrieval and processing, ensuring hallucination-free data handling and trustworthy outputs. The architecture is not just theoretical but has already been largely implemented and is actively in use within the German project SMARTA (Student Motivation and Reflective Training AI-Assistants).

Keywords:

AI, chatbots, assistants, architecture, trust, AI act, algorithm, academia, algorithmic prompt composer, qualified data retriever, SMARTA, trust-driven AI, Higher education.